<p>Chloride ingress is a major cause of corrosion in reinforced concrete, threatening long-term durability and structural integrity. This study presents an interpretable framework integrating Gaussian process regression with reinforcement learning-based hyperparameter optimization to predict chloride diffusion coefficient and electric charge passed. The model was trained on 132 concrete mixtures with varying proportions of cement, slag, fly ash, silica fume, water, and aggregates. A reinforcement learning policy network was employed to autonomously optimize the kernel parameters, leading to substantial performance improvements, as confirmed by ablation analysis. Compared with four benchmark models optimized using the same reinforcement learning procedure, including artificial neural network, gradient boosting, light gradient boosting machine, and adaptive boosting, the proposed framework demonstrated superior predictive accuracy and generalization. Robustness analysis under ± 10% input noise showed that over 92% of chloride diffusion predictions remained within ± 3 × 10⁻¹² m²/s. External validation using an independent dataset of 90 mixtures, produced under different conditions and material sources, further confirmed the model’s transferability. For practical implementation, the pretrained model was deployed as a user-friendly web platform for rapid durability assessment based on mix design parameters. Interpretability was achieved using explainable artificial intelligence analyses, including Shapley additive explanations, accumulated local effects, and partial dependence analysis. The results revealed mechanism-consistent insights showing that supplementary cementitious materials, particularly slag and silica fume, not only reduce chloride ingress but also mitigate the adverse effect of high water-to-binder ratios. Overall, the proposed framework provides a reliable and interpretable tool for performance-based durable concrete design.</p>